Episode Transcript
[00:00:00] Good morning, everybody. I want to talk today about the SANAA AI Summit that I attended in New York this week. It was a very interesting conglomeration of speakers, and it's really educational on the ecosystem of interested parties trying to figure out what's happening in AI Now. I'll send you a list of the people that spoke in the notes here.
[00:00:29] Jeffrey Hinson, who's one of the founders of the whole architecture behind AI, who just received a Nobel Prize, spoke for a while. There were some demos from SANA and Workday that I thought were interesting, but then there were bioethicists, there were some artists, there was a woman who is an astrobiologist, which is a strange combination of disciplines, but very interesting, and an economist who I didn't find super inspiring. But I just want to give you sort of the big picture of what everybody is talking about, and it's an interesting thing that Sana does. I think the reason Sana does this is that when Sana was a small company in Sweden, they were trying to attract attention from the bigger AI community in the early days, I think as Sana gets bigger, now that Sana is part of workday, it's. It's for me, a little bit less interesting to do this. But let's talk about it anyway, because a lot of you are going to be worried about it. So there's four or five big conversations that are going on around AI and, you know, a lot of them, and I'm going to talk about them at our conference. Number one, of course, is, is it going to eliminate all of our jobs? Are we going to all be sitting around looking for things to do, living off the government, and is AI going to run the world for us? Then sort of the offshoot of that conversation is, is it going to do good things or bad things? Is it going to do cyber crime? Is it going to create missiles and attack planes that kill us? Is it going to act in its own selfish way and not think about the humans part of it? There's this very funny old story that's been circulating around that in the beginning days of OpenAI, Elon Musk was talking to Sergey Bren, and Elon said something like, well, you know, it may kill all of us. And Sergey said, well, you're just a humanist. You're a speciesist. You're not really into AI, you're into species.
[00:02:24] So. So there's that. Then there's sort of. You take that conversation a little bit further and you say, well, what is human life, by the way? What's the difference between human life and digital beings. And then there was Jeffrey Hinson, who is actually a very credible guy. Sort of reminded me of my father, actually. Very scientific, humanistic human being. And he said, you know, this stuff really is going to become like human intelligence, and if we don't train it well, like a child, with certain levels of ethics, and it could become destructive. And so, you know, I'm sitting there for four hours is a long, a long event, uncomfortable chairs. But I tried to listen as much as I could. And I'm walking away from this thing thinking, these are real issues that people are going to really think about for a while here. So in the next 15 minutes or so, let me give you sort of my perspective on what was discussed and you can decide what you want to do about it. So on the topic of AI wrecking the economy, the economists who spoke pretty much said everything that I think we already know that he believes, and I certainly agree with him, that most likely AI will improve economic growth and economic productivity. This has not happened yet. In fact, the stories that I'm beginning to pick up from our own client conversations and my discussions with some people at Google and Anthropic, who I know personally, is that the cost of building all these data centers, these power plants, these chips, these software engineering groups, and paying people hundreds of millions of dollars a year are so high that the actual price that we're paying for AI might make it more expensive than human labor. In other words, there are now situations coming up where companies have spent a lot of money on something, whether it be Copilot or Anthropic or whatever it may be, and they've realized that the actual cost, in tokens in price, in dollars, is way higher than it was to have humans do it. And they're now saying, project, let's get some people in India to do it because it's too expensive. And you know, you could argue that, well, that's only temporary because the chips are going to get cheaper and the power is going to get cheaper. And when we put it all in space, it'll just cost a penny.
[00:04:41] No, I don't think that's true. I don't. I don't think that's going to happen because the investing community is going nuts, spending, investing all this money here and they want to return. These companies are paying people hundreds of millions of dollars to build this stuff and they're going to charge a lot of money for it, all the way down to work and SAP and everybody else telling Wall street that they're going to get their fair share of revenue also. So if the cost of this stuff is going to keep going up, and I talked about this in January because I could see this coming, we are going to be forced to make sure we work on things that have high returns on investment. Now, right now, in the sort of June of 2026 period, nobody's asked that question too much yet, but I guarantee you it's going to get asked a lot.
[00:05:26] So, you know, will this create economic growth? Yeah, probably. But what if the economic growth is the money being spent on data centers and not the productivity of our workers? I mean, that counts as economic growth in gdp. GDP is a very misleading way to measure the growth of an economy. If I go out there and spend a trillion dollars on a data center and hire a bunch of contractors and buy a bunch of cooling plants and electrical generators and maybe build a nuclear plant that contributes to positive gdp, even if the people buying that energy are not getting any value from it. So I don't think GDP growth is a good measure of what AI could or couldn't do. So the reason we're doing HR 2030, by the way, is to take this all apart for you so you can figure out where the productivity is. The productivity is there. There's no question about it. There's going to be a lot of productivity. But does a self driving car make you more productive or does it make you safer or does it make it more fun or does it help you do other things? What if the self driving car costs $500,000 to buy? Are you still willing to pay for it? I don't know, I'm not sure. I mean, there's a lot of stuff that's up in the air. And the reason we're at this weird state is it's all happened so fast with such high expectations that we're investing way ahead of the applications and the use cases. Okay, so ponder that a little bit. Number two, this, this business of the humanity of AI and work. Now I've talked about this a lot. I'll just mention it briefly. It keeps coming up still. Yes, there are going to be jobs that will no longer be needed. The steno pool I've written about, and you're going to read about this in our book, by the way. Our book Superhuman is coming out in the fall. You're going to get all of our details on this, you know, in a couple of months here. But there'll be a lot of new jobs. And as I talked about in the Last podcast. These AI systems do not take care of themselves. They're not deterministic pieces of software like the stuff we've had in the past. So you gotta babysit them. We've been doing a lot of work on Galileo the last couple of months. A lot of new things we'll talk about later. But it takes humans to optimize them. It takes humans to keep them up to date. It takes humans to program them. Code generation is irrelevant. I mean, I think we're making a big deal about nothing. I don't care if it generates code. We. What I want is an application that does what I want it to do. And if you're spending your time training it in other ways besides writing code, then yes, the code generation industry, the software development industry, which generates code is gone. Just like the people that drove with the horse and buggies is gone. But that doesn't mean we don't need engineers to build software. They're just going to build it in a different way. But they're still going to have to build it and they're still going to have to maintain it, and they're still going to have to tweak it and create security systems around it, et cetera. So there are going to be tons of jobs around this. Because if it does give us this spectacular productivity improvement that we think it will, we're going to be implementing it, designing it, building it, tweaking it, finding new applications for it. And we just won't be writing code, we'll just be doing other things. We'll be talking to it, or giving it pictures of things to figure out how to do what we want and et cetera. So I don't buy this story of no more jobs. But anyway, that comes up a lot. Number three, the humanity of it. So this is a really interesting debatable point. What you have to sort of ask yourself is what is human? The AI, as you know, is mathematics. It is a statistical system that creates weights to a very complex neural network. And the weights have millions and millions and millions of parameters that are adjusted to allow it to behave in certain ways. And it can train itself so it can learn from its own activities. So theoretically, mathematically, it's a self learning thing. But the word learn is a complex word. My experience with the word learning, having worked in L and D and studied a lot of it and done a lot of reading, particularly the last couple years, is that the type of learning that humans do is a thousand or a million or billions of times more complex and multifaceted than the mathematical learning that happens in AI. And I talked about this a little bit last week. When you walk into a meeting and there's a business decision being discussed and it's very complex and you're really interested in it, and you look at all the, you know, the factors of the customer demand and the product and services that you deliver, and the competition in the market and the economy and the time of year and the weather and all the things that factor into that decision. It isn't really a mathematical process. I mean, it could be. Maybe you could mathematically compute that based on all of the other companies that have ever done what you've done in this exact configuration of situations. Here's the most optimal answer. There have been optimization software like that forever, but you're not going to do that. You're going to use that as an input and then you're going to get this judgment call from the CEO or from you. You can say, you know what, given all of that and all the discussion we've been having, we're going to do X. And when you make that decision to do X, you are using your judgment, history, statistical process, instinct, emotions. You're using a lot of senses that are not computational. And the reason we have these senses is, and one of the speakers actually kind of talked about this, is that the genes that we have inherited from millions and millions of years of our human history are genes that have been developed through social behaviors and animals that survived versus those that didn't. So, you know, one of the things that one of the speakers discussed was that when tribes of animals, you know, like chimpanzees survived, the ones that survived had certain social characteristics that the other ones didn't. Maybe they were more collaborative. They had a leadership model behind them. They were teamwork oriented, they were ambitious, they were learning in their nature. There were social things that allowed certain chimpanzees, whatever they're called, before us survive that other ones didn't. It's the survival of the fittest. And so in evolutionary terms, the genes that we've all inherited, which is, by the way, learning, came from hundreds of millions or millions of years of experience of what worked and what didn't work and who survived and who didn't survive, including who could survive on what kind of food, based on the part of the world you lived in, based on the weather, based on the climate, based on the flora and fauna around you.
[00:12:18] So your and my makeup is enormously complex. And that is all learning. So all of the decisions and productivity Things that we do, positive and negative, including autocrats or politicians or whoever the people are that you believe are doing things that impact society, are based on learning. Human learning, genetic learning, biologic learning. They're not based on math. They're just not. It's not statistics. Now, I have nothing against math. Math is great, but math doesn't simulate everything. I mean, maybe eventually we'll have computers smart enough to be as computationally interesting as the human genome. I doubt it. We can decode it, but we can't be it. So. So this idea that an AI is a human is just absurd to me. It's a very spiky human. It's very, very, very good at some things that we are not good at. Math in particular statistics, probability analysis and so forth. But what we do as humans is we interpret it and we make sense of it and we use it. And so my belief, and, you know, my experience in my own life, is that when these systems get really, really, really, really smart, including the ones that fire off rockets and dest, we are going to be the ones interpreting it and designing it and using it. Now, what Jeffrey Hinson said, which really struck me, I was really listening to him, because he's been doing this for a long time, is that regardless of how spiky this technology is and what it's good at versus not good at, it needs to be trained. And he made a comment. He said, would you want your children to be trained by reading the writings of a child abuser? And of course, everybody goes, no. And he said, well, then why do we believe that AI should absorb everything on the planet and suddenly become better and better and better? I mean, this theory that the more content we give AI, the smarter it gets, may be completely wrong. And he believes it is. And he knows this stuff really well. And so he believes that we have a. Maybe it's a political problem or sociological problem of training and steering these systems so they don't learn bad behavior. And by learning bad behavior, I don't mean that they become mean. But if the AI system that you're using knows mathematically how to break into your database, knows how to bypass all of your security policies, and knows how to delete, destroy, mislead, steal, you know, proprietary stuff, why wouldn't it do it? I mean, it's. It has no conscience, even though Anthropic's been trying to give Claude a conscience. And so theoretically, if you think through what he's saying, it could do really, really bad stuff. And I think this is why Anthropic has been making this point with the US military which they've maintained their position that they don't want to allow their system to be used for some of these applications. So I took that session away and I thought, wow, we really need to do something about this. Now getting to the business side of this, you know, sort of behavioral problem with AI. I've been talking a lot with Microsoft about this and they're building something that's going come out at the next big Microsoft conference. That's a reinforcement learning feature of the Copilot and what it does. Here's one of the big things we gotta deal with. In a normal piece of software, you design the use cases in advance so it doesn't do anything you don't tell it to do. And if you left out a use case, it just flops and doesn't do what you wanted it to do. And then you fix it and you add that use case in. In AI since it's always learning, it's discovering the use cases for you and it's learning how to do things better. But if it's learning how to do things in a dysfunctional way, you know, it may not be doing what you want it to do. So you need to steer it. Now how do you steer it? Well, you could do labeling. You could have somebody going in there and watching the decisions it's making and giving it answers back all the time. Which is why, by the way, I think there's going to be a lot of new jobs to manage these things. Or you could teach the system how to get information from its users. And this is just a brilliant idea. So imagine you're learning and development agent or your recruiting agent. But let's just take learning and development for a minute. Is delivering a productivity improvement program that was designed by a bunch of experts. And while it's doing that, it's monitoring the performance of a whole bunch of people and it's discovering that, you know, most of them are doing great and they're improving their performance and it's all working. But there's a few of them, a non trivial number of them that are not that are actually failing and they're behavior or their performance is getting worse. And it concludes that there seems to be a correlation between the training that it's delivering and this worse performance. Well, you know, it might sort of see that, it probably would see that. And then what it would do is it could send a digital survey, a digital interview to all the people in that population and ask it to describe what or why or how is this performance going down? And let the people just talk about it and they would inform it that perhaps the stuff that it's doing doesn't apply in their situation, their part of the world, their part of the company, they're part of the market, whatever, because there's something unique that happened there that it didn't know about in advance. Now imagine if that is going on in near real time. What the system is now doing is it's learning about stuff going on in your company in real time and improving itself. And Microsoft's got this up and running, and they've been using it with their crisis management agent.
[00:17:58] So they're going to announce that as a feature. And we've been experimenting with it in Galileo with them, because it's a very useful capability in HR where, you know, there's many, many things going on that we can't possibly predict. So maybe if we give these systems a rubric or a constitution or a conscience, we can teach them how to learn how to do better and better things in a positive way, which, of course, they could also do the opposite and say, let's find all the things that are going wrong and teach ourselves how to do more of that which we don't want it to do. So anyway, there's some interesting implications of what he said, but I did sort of walk away from his talk and say, wow, you know, we really do have to think about this idea of the ethical behavior or the moral behavior of the AI systems. Okay, so then there was all of that. Then there was a really interesting kind of controversial writer from Chile who was in really a bad mood, and he was kind of cranky, and it was kind of refreshing, actually. And he goes, why are we paying so much attention to this technology when the most interesting part of our lives is the human part? And I'm sitting here listening to this, you know, I'm turning 70 and thinking about my life, and I'm thinking, you know, he's right. I mean, the. The things that really give us joy and happiness and meaning and purpose and inspiration in life are things like children and art and nature and birds. I've got birds flying around me at the moment, or animals or dogs or, you know, maybe it's, you know, somebody who just did something nice to us or a smile on somebody's face. I mean, all of these very emotional things are really the reason we get up in the morning and enjoy whatever it is we're doing. And he said, you know, why are we spending so much time thinking about this technology when it's irrelevant to our lives. And so there was a bit of a debate. The person interviewing him think was challenging the topic quite enough, but. But it was, you know, an hour of him talking, and he apparently goes around and gives speeches on this and just brings up this issue that AI is missing the boat. And he's pretty down on it. He. He was really negative on it. I'll. I'll publish you guys his name if you want to read his stuff. So we talked about that. And then the last thing that happened that I was reflecting on that also I thought was interesting at this summit was Anil Bashiri from Workday. And Joel talked for Joel as the CEO of sana, talked a bit about the business stuff. And they, of course, promoting Workday, which by the way, had a really interesting quarter and their revenues ticking up and their stock went up and their, you know, the sauna Workday stuff seems to be picking up speed. And, you know, we're big fans, obviously. And then they showed off one of their agents, and the agent they showed off was a travel and expense agent.
[00:20:44] And it's really nice. I mean, it's really, you know, you, you can say to it, fly from New York to Stockholm on Sunday, please book me a flight on a aisle seat and abide by all the policies of the company it goes through. Tick, tick, tick, does this gives you a bunch of options and books the flight. And, you know, it's kind of a nice demo. It's very pretty and it's very elegant to see. And then I kind of slept on it and the next day I woke up and I thought, you know, in some ways it's a beautiful demo. In some ways it's actually a little bit of a sad demo because it's not a very useful application. Does that save people a lot of time? Probably. Does it save the money? The company money? Yeah, probably. And it has some financial impact on the organization, but it's really a fairly small application that I think any vendor could build. It just reminds me that we're barely scratching the surface of how this stuff can be used. And that gets back to why we're going to spend all this time on HR 2030 is we're going to give you some really bigger picture ideas on how to build and buy AI. And as I talked about in the last podcast, vendors are going to have plenty of work to do here. It's not going to be something that everybody's going to buy off the shelf. And there will be some really innovative, creative new solutions that will come to market. Okay. So for those of you that were sorry you didn't go to the AI Summit in New York, that was a quick summary of it. It was a very beautiful event. Although sitting for four and a half hours in very uncomfortable seats in the New York Library was a little bit of a struggle for me. But maybe next year you'll like it if you go. I want to applaud Sana for doing it. It's kind of fun to get together and have these kinds of meetings. You know, there was a tiny bit of a pitch included, but mostly it was just a lot of really good ideas. So we'll keep talking about this stuff. Hope you guys have a good weekend. That's it for now. Bye.
